While many AI projects are competing over parameter counts, GPU numbers, and model size, this demo from @Acurast Acurast takes a different approach: it lets an Android phone decide where the next Tetris piece should go.

At first glance, this looks like a pretty casual little game. But what I find interesting isn’t that “AI can play Tetris.” That’s old news. What’s really worth looking at is how it breaks an AI task down in a different way.

The game first generates five possible placements based on the current board, explaining the likely outcome of each choice—for example, whether it could clear a line or leave a gap. This information is then sent via an HTTP request to a phone in the Acurast network. A Laya model running on the phone’s CPU picks an option, and the result is sent back to the game.

The whole process can be summed up in one sentence: software generates the options, AI makes the decision, and phones perform the computation.

Laya is not the kind of large language model we commonly see in chatbots. It’s a System One decision model: it doesn’t write long-form text or perform complex reasoning. Instead, it quickly chooses from a set of options that have already been provided and returns the probability associated with each one.

According to data published by Acurast, these kinds of short decisions typically take around 0.2–1 second on a phone’s CPU. In addition to Tetris, the team has created nine other demos, including Snake, Doom, email classification, chat moderation, and prompt injection detection. Each response comes from a real phone on the network. The page also shows the device’s country, chip model, and number of decisions processed, and provides on-chain records for users to inspect.

I think this approach is far more practical than “making phones run a huge model on their own.”

Real-world AI tasks don’t all require generating thousands of words. A wallet checking whether an authorization complies with a user’s rules, a transaction aggregator choosing the cheaper of several routes, or an on-chain agent deciding which action to take next—these tasks often have a clearly defined set of options.

Software can first calculate all the valid options, then a small model can choose among them based on rules set by the user. The final result can then be checked independently. This means AI doesn’t have to do all the computation, and it can’t arbitrarily take actions outside the set of options.

Of course, these models are nowhere near万能.

In Acurast’s email classification test, Laya correctly classified 144 out of 160 emails, for an accuracy rate of about 90%. The news classification demo can tell whether a headline looks like clickbait, but it can’t determine whether the content is true. When chats move too quickly, the moderation queue also starts to grow. The probabilities returned by the model shouldn’t be treated directly as reliable “confidence scores,” either.

So it may be better suited to serving as a fast, low-cost first-layer filter, rather than making the final decision on someone’s behalf.

The current reference cost published by the team is less than 1 $ACU per day to run your own Laya instance. Actual costs and availability still depend on the configuration and network conditions. The figure itself may not be the most important thing; what’s more noteworthy$ACU is that this refers not to an abstract narrative, but to a real computing service that can be called, billed, and verified.

The next phase of AI may not simply be about training ever-larger models. Many needs that make their way into real products may involve nothing more than a classification, a filter, or a small decision.

If these unassuming but frequent tasks can be handled by phones distributed around the world, then what Acurast aims to offer is not just “cheap computing power,” but a distributed execution environment suited to lightweight AI decision-making.

If an idle phone could handle one small AI task for you, what would you want it to do first?

This article reflects my personal observations on the project’s technology and use cases and does not constitute investment advice.